From 3cd3473bf7a3b41484baa86d9092248d78e7af39 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 21 四月 2023 17:17:37 +0800
Subject: [PATCH] docs

---
 funasr/train/trainer.py |   34 ++++++++++++++++++++++++++++++----
 1 files changed, 30 insertions(+), 4 deletions(-)

diff --git a/funasr/train/trainer.py b/funasr/train/trainer.py
index efe2009..7c187e9 100644
--- a/funasr/train/trainer.py
+++ b/funasr/train/trainer.py
@@ -94,7 +94,7 @@
     wandb_model_log_interval: int
     use_pai: bool
     oss_bucket: Union[oss2.Bucket, None]
-
+    batch_interval: int
 
 class Trainer:
     """Trainer having a optimizer.
@@ -186,7 +186,7 @@
                 logging.warning("No keep_nbest_models is given. Change to [1]")
                 trainer_options.keep_nbest_models = [1]
             keep_nbest_models = trainer_options.keep_nbest_models
-
+ 
         output_dir = Path(trainer_options.output_dir)
         reporter = Reporter()
         if trainer_options.use_amp:
@@ -560,12 +560,38 @@
         # [For distributed] Because iteration counts are not always equals between
         # processes, send stop-flag to the other processes if iterator is finished
         iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")
-
+        
+        #get the rank
+        rank = distributed_option.dist_rank
+        #get the num batch updates
+        num_batch_updates = 0
+        #ouput dir
+        output_dir = Path(options.output_dir)
+        #batch interval
+        batch_interval = options.batch_interval
+ 
         start_time = time.perf_counter()
         for iiter, (_, batch) in enumerate(
             reporter.measure_iter_time(iterator, "iter_time"), 1
         ):
             assert isinstance(batch, dict), type(batch)
+
+            if batch_interval > 0 and (not distributed_option.distributed or rank == 0):
+                if hasattr(model, "num_updates") or (hasattr(model, "module") and hasattr(model.module, "num_updates")):
+                    num_batch_updates = model.get_num_updates() if hasattr(model,"num_updates") else model.module.get_num_updates()
+                if num_batch_updates % batch_interval == 0:
+                    if options.use_pai and options.oss_bucket is not None:
+                        buffer = BytesIO()
+                        if hasattr(model, "module"):
+                            torch.save(model.module.state_dict(), buffer)
+                        else:
+                            torch.save(model.state_dict(), buffer)
+                        options.oss_bucket.put_object(os.path.join(output_dir, f"{num_batch_updates}step.pb"), buffer.getvalue())
+                    else:
+                        if hasattr(model, "module"):
+                            torch.save(model.module.state_dict(), os.path.join(output_dir, f"{num_batch_updates}step.pb"))
+                        else:
+                            torch.save(model.state_dict(), os.path.join(output_dir, f"{num_batch_updates}step.pb"))
 
             if distributed:
                 torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
@@ -811,4 +837,4 @@
         else:
             if distributed:
                 iterator_stop.fill_(1)
-                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
\ No newline at end of file
+                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)

--
Gitblit v1.9.1